What to expect at GTC 2025

AI-powered pose detection takes center stage at NVIDIA GTC 2025

Imagine tracking a golfer’s swing or a sprinter’s stride in real time—artificial intelligence (AI) is revolutionizing sports performance, and Latent AI is bringing that power to NVIDIA GTC, March 17–21, 2025. Join us and our partner Dell Technologies at Booth #1423 for our “Strike AI Pose” game—test your reflexes with real-time pose detection. The top three scorers win a Dell Precision 5490 14″ laptop! 

What’s unique about Latent AI’s “Strike AI Pose”?

Our demo, powered by NVIDIA RTX 6000 Ada GPUs on Dell workstations, runs three optimized AI models simultaneously. It delivers a 3x speedup for smooth, real-time pose detection. 

1. Depth Estimation – We sharpen player focus with MiDaS, leveraging a SWIN Transformer for robust monocular estimation.
2. Object Detection – We identify players reliably, even in crowded expo halls, using YOLOv8s.
3. Pose Estimation – YOLOv8s-Pose tracks key body points to analyze and align the participant’s pose with the reference.

Real-time movement tracking demands exceptional precision, where even the smallest delay can affect accuracy. We can deliver a 3x speedup for smooth, real-time pose detection with LEIP Optimize. LEIP Optimize provides hardware-aware quantization and compilation that optimizes AI models to interact better with your device’s available hardware, boosting inference speeds and minimizing latency. Without optimization, these tasks would either require immense processing power or suffer from performance bottlenecks. The table below illustrates LEIP’s optimization improvements measured in FPS (frames per second) to describe how quickly the AI model processes input data and produces results. Reflecting the system’s responsiveness and efficiency—the higher the FPS, the smoother and more responsive the AI experience.

How the technology enables real-world use cases

Beyond gaming, AI is shaping the future of team sports analytics, e-sports performance enhancement, and personalized fitness training. Pose detection—tracking an athlete’s body position in real time—helps coaches refine technique, spot fatigue, or analyze movements like a golfer’s swing or a sprinter’s stride. The challenge? It must be precise, quick, and workable on the gear teams already use, like on cameras or embedded in other edge devices, without breaking the bank or slowing things down.

Instead of one big, clunky AI model trying to do it all, splitting the job across specialized models makes AI applications smarter and leaner. As with the Strike AI Pose game, each model serves a specific purpose. In a golf scenario, one model might focus on spotting key joints (like hips, knees, or elbows), another on tracking motion paths, and a third on interpreting what it means (e.g., “that swing needs more hip rotation”). This divide-and-conquer approach lets each model run efficiently, tailored to its task, so you get pinpoint accuracy without overloading the system. For sports businesses, it means better insights—like catching a swimmer’s inefficient arm pull—without needing supercomputers.

More at GTC 2025

At Booth #1423, explore how our Latent AI Efficient Inference Platform (LEIP) powers low-energy AI on tough edge devices. We’ll be showing how:

• The LEIP SDK enables you to design, optimize, and deploy models directly into your app. It automates model-hardware optimization for rapid prototyping and deployment, streamlines the entire ML pipeline for speed, consistency, and scalability, and securely encrypts multiple models for one or more hardware targets.
• LEIP Design allows you to build, optimize, and deploy an AI model tailored to your specific hardware. LEIP Design streamlines the typically complex and disjointed process of AI model development, making it both accessible and efficient.
• LEIP Optimize transforms a radio frequency (RF) signal classification model into an edge-ready solution. On the Jetson Orin AGX (64GB, Jetpack 5.1.3), it delivers a 3x GPU RAM reduction, 11.7x inference rate boost, and 10x energy efficiency improvement—all while preserving accuracy.

Visit our event page for more information on optimizing your models for speed and size—unlocking on-device magic for your apps. Don’t miss it!